In this paper, we have aimed to develop a system that will help waste
collectors segregate different types of waste without needing much human intervention.
We have experimented with various deep learning and transfer learning techniques to
determine which model is more suited for this purpose. The dataset we used contained
8369 images that are classified into 9 classes: batteries, clothes, e-waste, glass, light
bulbs, metal, organic, paper, and plastic. We used models like VGG16, Inceptionv3,
ResNet50, MobileNET, NASNetMobile and Xception. We have also conducted a
survey to know about the waste management habits of the respondents. Our
experiments showed that models like MobileNET gave us the best accuracy of 93.17%
and identified all the waste categories correctly and the Xception model predicted
images correctly with the use of both Adam and Adadelta.
Keywords: Deep learning and transfer learning, Inception v3, MobileNET, NASNetMobile, ResNet50, VGG16, Xception.
